Spaces:
Running on Zero
Running on Zero
Commit History
4 days ago
-
fix: an adapter whose weight pairs are named lora_down/lora_up is refused at Validate, naming both namings and that this model's loader reads only lora_A/lora_B, instead of passing every check and failing inside the run with "list index out of range"
Unverified signaturedaa617a -
fix: an adapter whose tensor data is cut short or out of place is refused at Validate, before any GPU time
Unverified signatureinspect_lora now applies the safetensors loader's own layout rules from the header alone: in offset order every tensor starts where the previous one ended (the first at 0) and spans exactly its shape times its dtype size, and 8 + header + data must be the file's size. A file shorter than that, which is what an interrupted download leaves under its final name, is refused as truncated, naming the size the header promised and the size on disk; gaps, overlaps, wrong spans and bytes after the last tensor are format errors naming the tensor. These files used to pass every stage and fail only in the worker, inside load_lora_weights, after the GPU reservation, the enhancer and the condition prep. The tests check that the real safetensors library refuses every layout refused here; all 20 real adapter headers the recon saved (the official 8.9 GB distillation adapter included) still pass, and each one's header-derived size equals its Hub-listed size.95c2c72
5 days ago
-
fix: the LoRA validator checks shape fit with no weights, and tells the truth before a GPU is reserved
Unverified signatureStage 6.5 (does each low-rank factor fit its target module) only ran when a live transformer was passed. The UI never passes one, and Generate skips validation for an entry that already carries a report, so a UI user never got it. It was also wrong for the most common key convention: it stripped diffusion_model. without diffusers' native-to-diffusers rename, so the official LTX-2.5 adapter was told 26 of its 1660 modules were absent, "a normal version difference". - lora/transformer_linear_shapes.json lists the 1660 Linear weights of the pinned transformer (Lightricks/LTX-2.5-Diffusers @871165de), derived from the four shard headers, with that revision and a sha256 of its canonical JSON. `python -m ltx25_studio.lora.manifest` re-derives it from the Hub with header range reads only; an opt-in network test asserts equality. - lora/keys.py resolves adapter keys the way diffusers will: its own _convert_non_diffusers_ltx2_lora_to_diffusers where importable, the same rename as a…9d84791
-
fix: the pipeline takes PipelineQuantizationConfig, not BitsAndBytesConfig
Unverified signatureThe Space deploy failed on a type, not on disk and not on RAM: ValueError: `quantization_config` must be an instance of `PipelineQuantizationConfig`. A bare BitsAndBytesConfig is what a COMPONENT's from_pretrained takes; DiffusionPipeline.from_pretrained rejects it. Read at the pinned SHA rather than guessed: quant_backend="bitsandbytes_8bit" is a key in both backend mappings, quantizers/auto.py:45 and :57. This is a rung past the one the last test was built for. reached -> correct argument -> ARGUMENT OF THE RIGHT TYPE FOR THE CALLEE. The kwargs test proved the config arrived and could not tell the two classes apart, because both arrive equally well. Bound B is now an ARGUMENT -- components_to_quantize=["transformer"] -- rather than an intent a source scan had to assert. That is strictly better: widening is a change to a value a test reads, not a change to what a call happens to produce. So test_only_the_transformer_is_quantised is retired, with its reason kept where it lived…b55f9af
-
test: spec validator stages 5.5 and 6.5, and the revised stage 4
Follows the PLAN amendment. Stage 7 is necessary but not sufficient: an LTX-Video-2B adapter whose module names match ours attaches to >0 modules and passes it, and only a per-module dimension comparison catches it. The fixtures now carry real dimensions, including the two-stream shape (video 4096, audio 2048, 4x feed-forward expansions), so the spec can require that a 2048 tensor at an audio module is accepted while a 2048 tensor at a video module is rejected. A rule that just looks for 2048 would reject the most-downloaded LTX LoRA on the Hub. Also relaxes the hardcoded-filename sweep: it now flags string literals that ARE filenames, not prose in the UI layer that mentions the extension.d86776a -
test: spec the mode registry, the never-empty-output invariant, app build and layering
The generate.py specs are the ones that matter: failure is injected at four depths (model load, decode, missing output file, zero-byte output file) and each must surface a distinct typed error rather than a falsy path. Also asserts there is no newest-file-in-output-dir fallback, statically and behaviourally — the 2.3 version had no run correlation and would show a previous run's video for a run that failed. Mode identifiers are fixed here as the contract: PLAN §3.2 numbers and labels the modes but never names them.1e0df90 -
test: spec errors, constraints, model_files and the LoRA validator
Red-first specs for PLAN §4.3 (7-stage LoRA validator), §5.2 (single source of truth for Hub coordinates), §5.5 (frame/dimension constraints) and the error hierarchy. Runs with no GPU, no network and no model weights: safetensors fixtures are synthesised byte-for-byte in pure Python, which is also the only way to craft the truncated and corrupt-header files stage 3 has to distinguish.5861856